Offshore wind speed prediction method based on DAI-LSTM-AT model

Abstract

The invention discloses an offshore wind speed prediction method based on a DAI-LSTM-AT model, and relates to the technical field of meteorological prediction and artificial intelligence, and the method comprises the steps: obtaining meteorological grid reanalysis data of a target region and buoy observation data of a target point position; performing height conversion and time alignment processing on the buoy observation data according to the meteorological grid reanalysis data, and performing normalization processing on the meteorological grid reanalysis data and the processed buoy observation data to obtain normalized data; and constructing an offshore wind speed prediction model according to the long and short-term memory network, a dynamic attention interpolation method and a Transform structure, and inputting normalized data to obtain an offshore wind speed prediction value of a target point position. According to the method, the problems of insufficient data precision of observation sparse data areas, limited precision of deep and far sea reanalysis data, weak generalization and low prediction precision of a traditional method are solved, and accurate and efficient offshore wind speed prediction of accurate positions is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of weather prediction and artificial intelligence, in particular to a sea wind speed prediction method based on a DAI-LSTM-AT model. BACKGROUND

[0002] Under the background of the growing demand for renewable energy worldwide, as an important renewable energy source, wind energy has attracted more and more attention from countries and regions due to its abundant yield, mature technology and pollution-free characteristics.

[0003] The spatiotemporal prediction of offshore wind farms is a key challenge in environmental monitoring and renewable energy systems. As an unstable power source, wind energy has the characteristic of random fluctuations, which brings many difficulties to its prediction.

[0004] As an unstable power source, offshore wind power also has the characteristic of random fluctuations in wind speed, which brings many difficulties to wind speed prediction and wind power prediction. Compared with onshore wind power, offshore wind power has more abundant wind resources, greater wind power potential, higher average wind speed, allows installation of larger wind turbines and smaller environmental impact. However, the complexity and variability of the marine environment make it difficult to accurately predict offshore wind speed, and inaccurate prediction will cause damage to wind power equipment, disrupt the balance of power supply and demand, and prevent ocean-going vessels from determining whether they can go to sea.

[0005] Therefore, there is an urgent need for an accurate and efficient offshore wind speed prediction method for a specific location. SUMMARY

[0006] The purpose of the present application is to provide a sea wind speed prediction method based on a DAI-LSTM-AT model, to solve the problems of insufficient data accuracy in sparse data observation areas, limited data accuracy in deep sea reanalysis, weak generalization and low prediction accuracy of traditional methods, and to achieve accurate and efficient sea wind speed prediction for a specific location.

[0007] To achieve the above purpose, the present application provides a sea wind speed prediction method based on a DAI-LSTM-AT model, comprising the following steps: S1, obtaining meteorological grid reanalysis data of a target area and buoy observation data of a target point position; S2, according to the meteorological grid reanalysis data, performing height conversion and time alignment processing on the buoy observation data, and performing normalization processing on the meteorological grid reanalysis data and the processed buoy observation data to obtain normalized data; S3, constructing a sea wind speed prediction model according to a long short-term memory network, a dynamic attention interpolation method and a Transformer structure; S4, input the normalized data into the offshore wind speed prediction model to obtain an offshore wind speed prediction value of the target point position.

[0008] Preferably, the specific content of constructing the offshore wind speed prediction model in S3 according to the long short-term memory network, the dynamic attention interpolation method and the Transformer structure comprises: obtaining meteorological grid reanalysis historical data of a region, buoy observation historical data of a point position and offshore historical wind speed values of the point position; the point position is in the region; performing height conversion and time alignment processing on the buoy observation historical data according to the meteorological grid reanalysis historical data, and performing normalization processing on the meteorological grid reanalysis historical data and the processed buoy observation historical data to obtain normalized historical data; constructing a training set, a validation set and a test set according to the normalized historical data and the offshore historical wind speed values; constructing a convolution long short-term memory network; the convolution long short-term memory network is obtained by replacing matrix multiplication in the long short-term memory network with convolution operation; constructing a dynamic attention interpolation module based on a learnable self-attention mechanism; constructing an LSTM-Transformer module based on a Transformer structure; connecting the convolution long short-term memory network, the dynamic attention interpolation module and the LSTM-Transformer module to obtain an initial model of the offshore wind speed prediction model; training, validating and testing the initial model based on the training set, the validation set and the test set, and taking the trained, validated and tested initial model as the offshore wind speed prediction model.

[0009] Preferably, the convolution long short-term memory network comprises a forget gate, an input gate, an output gate and a memory cell; the multiplication used by the forget gate, the input gate and the output gate is convolution multiplication; the convolution long short-term memory network performs feature extraction on the normalized historical data to obtain spatio-temporal features.

[0010] Preferably, the specific content of constructing the dynamic attention interpolation module based on the learnable self-attention mechanism comprises: obtaining latitude and longitude according to the spatio-temporal features; projecting the latitude and longitude through a fully connected encoding layer to obtain a position query vector; mapping the spatio-temporal features through a linear projection layer to obtain spatio-temporal mapping features; the spatio-temporal mapping features have the same dimension as the position query vector; defining an embedding matrix and adding it to the spatio-temporal mapping features element by element to obtain a spatial position fusion feature; The similarity of the position query vector and the spatial position fusion feature is calculated, and a weighted sum is performed to obtain a global vector. The global vector is subjected to projection processing and layer normalization processing to obtain a position feature sequence, and a dynamic attention interpolation module is obtained.

[0011] Preferably, the specific content of the similarity calculation of the position query vector and the spatial position fusion feature and the weighted sum includes: The spatial position fusion feature is flattened to obtain a key vector and a value vector. The similarity of the query vector and the key vector is calculated to obtain an attention weight, and the value vector is weighted and summed through the attention weight to obtain a global vector.

[0012] Preferably, the Transformer structure used in the LSTM-Transformer module is obtained by adding a linear layer and a normalization layer in the initial Transformer Encoder structure.

[0013] Preferably, the loss function in the offshore wind speed prediction model updates the weight parameters through a back propagation algorithm; and the expression of the loss function is: ; ; ; wherein, is a mean square error, i is a data sample size, i =1,2,…, n , n is a positive integer, y i is i the real value of the wind speed at time t, y pred is a model prediction value, grad is the absolute error of the first derivative of the prediction sequence and the real sequence, y i+1 is i the real value of the wind speed at time t+1, is i the predicted value of the wind speed at time t+1, is i the predicted value of the wind speed at time t, loss is a loss function.

[0014] The application provides an offshore wind speed prediction system based on a DAI-LSTM-AT model, which is used to implement the offshore wind speed prediction method based on the DAI-LSTM-AT model, and comprises: The data acquisition module is configured to acquire meteorological grid reanalysis data of a target area and buoy observation data of a target point position. The data processing module is configured to perform height conversion and time alignment processing on the buoy observation data according to the meteorological grid reanalysis data, and perform normalization processing on the meteorological grid reanalysis data and the processed buoy observation data to obtain normalized data. The model construction module is configured to construct a sea wind speed prediction model according to a long short-term memory network, a dynamic attention interpolation method and a Transformer structure. The prediction module is configured to input the normalized data into the sea wind speed prediction model to obtain a sea wind speed prediction value of the target point position.

[0015] The application provides a computer device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor executes the computer program to implement the content of the sea wind speed prediction method based on the DAI-LSTM-AT model.

[0016] The application also provides a computer readable storage medium having a computer program stored thereon, wherein the computer program is executable by a processor to implement the content of the sea wind speed prediction method based on the DAI-LSTM-AT model.

[0017] In summary, the sea wind speed prediction method based on the DAI-LSTM-AT model provided by the application has the following beneficial effects compared with the traditional technology: under the framework of a deep learning model, the sea wind speed prediction model constructed based on a long short-term memory network, a dynamic attention interpolation method and a Transformer structure solves the problems of insufficient data precision in sparse data observation areas, limited data precision in deep sea reanalysis data, weak generalization and low prediction precision in the traditional method, and realizes accurate and efficient sea wind speed prediction for precise positions, which can be widely used in sea work ship route point wind speed prediction, storm route point wind speed accurate prediction, and also provides strong support for the implementation of sea wind farms, equipment maintenance and development of potential wind power regions, and accordingly reduces the investment of manpower and funds.

[0018] The technical method of the application will be further described in detail below with reference to the accompanying drawings and examples. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 The flowchart of the sea wind speed prediction method based on the DAI-LSTM-AT model of the application is shown in the figure. Figure 2 The framework diagram of the sea wind speed prediction method based on the DAI-LSTM-AT model of the application is shown in the figure. Figure 3This is a structural diagram of the long short-term memory network in this invention; Figure 4 This is a structural diagram of the convolutional long short-term memory network in this invention; Figure 5 This is a comparison diagram of multiple models of the No. 1 buoy of this invention; Figure 6 This is a comparison diagram of multiple models of the No. 2 buoy of this invention; Figure 7 The present invention provides a block diagram of a marine wind speed prediction system based on the DAI-LSTM-AT model. Detailed Implementation

[0020] The technical method of the present invention will be further described below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of this application.

[0021] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the scope of this application and its application or use.

[0022] Techniques, systems, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, they should be considered part of the instruction manual.

[0023] In all the examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0024] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0025] like Figure 1 Flowchart and Figure 2 As shown in the framework diagram, the method for predicting sea wind speed based on the DAI-LSTM-AT model includes: Step S1: Acquire meteorological grid reanalysis data for the target area and buoy observation data at the target location. The meteorological grid reanalysis data includes hourly wind speed, wind direction, sea surface pressure, and sea surface temperature; the buoy observation data includes wind measurement data used to verify the buoy.

[0026] Step S2: Based on the meteorological grid reanalysis data, perform altitude conversion and time alignment processing on the buoy observation data, and normalize the meteorological grid reanalysis data and the processed buoy observation data to obtain normalized data.

[0027] Specifically, step S2 is to first convert the buoy wind speed into the wind speed corresponding to the 10m elevation using the wind profile formula, that is, the height corresponding to the meteorological grid reanalysis data, and to reduce the sampling of the buoy wind speed and align it with the time stamp of the meteorological grid reanalysis data. Then, the meteorological grid reanalysis data and the processed buoy observation data are normalized. Taking the hourly sea wind speed, sea surface pressure and sea surface temperature data as an example, the average value and the standard deviation are obtained in the sea wind data at all time points, and the sea wind data is processed using the Z-Score normalization method, and the formula is as follows: ; Among them, X is the current data, X mean is the average value of the current data, X std is the standard deviation of the current data, X norm is the result of the current data after normalization processing.

[0028] Step S3, according to the long short-term memory network, the dynamic attention interpolation method and the Transformer structure, a sea wind speed prediction model is constructed.

[0029] Further, step S3 can be replaced by steps S301-S308: Step S301, the meteorological grid reanalysis historical data of the region, the buoy observation historical data of the point position and the sea historical wind speed value of the point position are obtained. The point position is in the region.

[0030] Step S302, according to the meteorological grid reanalysis historical data, the height conversion and time alignment processing of the buoy observation historical data are performed, and the meteorological grid reanalysis historical data and the processed buoy observation historical data are normalized to obtain normalized historical data.

[0031] Step S303, according to the normalized historical data and the sea historical wind speed value, a training set, a validation set and a test set are constructed.

[0032] In an exemplary embodiment of the present application, 24-hour hourly reanalysis data of wind speed, wind direction, sea surface pressure and sea surface temperature with a resolution of 0.25°x0.25° on the east coast of a certain country, and 24-hour historical data of the two national data buoy centers (NDBC) buoys in the region, i.e. buoy No. 44011 and buoy No. 41002, including wind speed data, are obtained. According to the ratio of 7:2:1, a training set, a validation set and a test set are constructed, and then input features and labels are obtained according to the set historical sliding window size and prediction window size (in this application, the historical data window is 6 and the prediction data window is 1 for single-step prediction data set making), which are named as X train , Y train , X val , Y val , X test and Y test , respectively. The input features and labels belonging to the training set are added to a new list in the form of a tuple, and then divided by batch_size (batch size / batch size). Batch_size (batch size / batch size) refers to the number of data samples processed simultaneously in one training iteration.

[0033] In step S304, a convolutional long short-term memory network is constructed. The convolutional long short-term memory network is obtained by replacing the matrix multiplication in the long short-term memory network with convolution operation, and includes a forgetting gate, an input gate, an output gate and a memory cell. The multiplication used by the forgetting gate, the input gate and the output gate is convolution multiplication. The convolutional long short-term memory network extracts features from the normalized historical data to obtain spatio-temporal features.

[0034] Figure 2 W u is the weight matrix of the update gate, X t is the time series data input at time t, X t-1 is the time series data input at time t-1, X t+1 is the time series data input at time t+1.

[0035] The traditional long short-term memory (LSTM) network structure layer is a recurrent neural network structure layer, and the initial purpose is to solve the long-term dependence of gradient disappearance or explosion of the typical recurrent cycle neural network (RNN) model. The specific structure is as follows Figure 3 ​As shown. LSTM neurons can add, maintain, or delete information transmitted through the feedback loop. The structure of an LSTM includes a forget gate, an input gate, an output gate, and a memory unit. Each gate consists of a sigmoid neural network layer that selectively allows some information to pass through based on the input value. This sigmoid layer outputs a value between 0 and 1 to indicate how much information it can allow to pass through, thus controlling how much information to retain based on new inputs and previous iteration values. The specific formula is as follows: Input Gate: ; in, i For input gate, t For time steps, i t In time step t The input gate activation vector, delta It is the Sigmoid activation function. W i This is the weight matrix. h t-1 This is the hidden state from the previous time step. x t The input vector at the current time step. b i This is a bias term.

[0036] Forgotten Gate: ; in, f For the Gate of Oblivion f t In time step t The forget gate activation vector, W f This is the weight matrix. b f This is a bias term.

[0037] Output gate: ; in, o For output gate, W o This is the weight matrix. b o This is a bias term.

[0038] Cellular state: ; in, C t In time step t The cell state vector, C t-1 In time stept -1 cell state vector In time step t The candidate cell state vector.

[0039] Memory: ; in, h t In time step t The hidden state vector, o t Let be the output gate activation vector, and tanh be the hyperbolic tangent function.

[0040] Candidate state: ; in, c As a candidate state, W c This is the weight matrix. b c This is a bias term.

[0041] Figure 3 In W u To update the gate weight matrix, X t Input the time series data at time t.

[0042] Convolutional Long Short-Term Memory (ConvLSTM) networks are structurally similar to LSTMs, but replace the matrix multiplication operations of LSTMs with convolution operations, enabling the simultaneous capture of spatiotemporal features. ConvLSTMs inherit the temporal memory advantage of traditional LSTMs while simultaneously endowing the model with local feature extraction capabilities through the spatial sliding of the convolution kernels. Specifically... Figure 4 As shown.

[0043] Step S305: Construct a dynamic attention interpolation module based on a learnable self-attention mechanism. Specifically, step S305 includes: The specific content of building a dynamic attention interpolation module based on a learnable self-attention mechanism includes: Latitude and longitude are obtained based on spatiotemporal characteristics.

[0044] The latitude and longitude are projected through a fully connected coding layer to obtain a location query vector.

[0045] Spatiotemporal features are mapped using a linear projection layer to obtain spatiotemporal mapped features. These spatiotemporal mapped features have the same dimension as the location query vector.

[0046] The embedding matrix is defined and added element by element with the space-time mapping feature to obtain a spatial position fusion feature.

[0047] Similarity calculation is performed on the position query vector and the spatial position fusion feature, and weighted summation is performed to obtain a global vector. Specifically, the spatial position fusion feature is flattened to obtain a key vector and a value vector, similarity calculation is performed on the query vector and the key vector to obtain an attention weight, and the value vector is weighted and summed through the attention weight to obtain the global vector.

[0048] The global vector is subjected to projection processing and layer normalization processing to obtain a position feature sequence, and a dynamic attention interpolation (DAI) module is obtained.

[0049] The DAI module is a custom deep learning layer designed for the fusion of meteorological grid data and point position observation data. The core function of this module is to replace the traditional and static grid index matching method, and automatically extract the most relevant spatial features from the input meteorological field through a dynamic and data-driven learnable self-attention mechanism.

[0050] The DAI module processing flow and internal structure are as follows: first, the DAI module receives the spatio-temporal features extracted by the previous convolutional long short-term memory network. The DAI module internally maintains two key trainable parameters, namely the latitude (buoy_lat) and longitude (buoy_lon) of the buoy, which are initialized as the real geographical coordinates of the buoy, but will be dynamically adjusted by the back propagation algorithm during the training process to find the optimal position at the feature level. Next, the two scalar parameters are projected into a high-dimensional representation space through a fully connected encoding layer to generate a query vector (Query) representing the target position query intent. At the same time, the input feature map is first uniformly mapped to the same model dimension (d_model) as the query vector through a linear projection layer to ensure spatial consistency. To give the grid space position information, the module defines a trainable position embedding matrix, which is added element-wise to the projected feature map, so that the features of each grid point are fused with their geographical position information. Subsequently, the spatial features fused with position information are flattened, ready to participate in attention calculation as keys (Key, K) and values (Value, V). Then, the learnable buoy position query vector is calculated for similarity with the flattened global spatial feature key vector. These weights essentially quantify the importance of each grid point feature in describing the wind speed at the target buoy under the current model conditions. The attention weights are then used to weight the sum of the value vector, finally outputting a context vector that aggregates the most relevant global information, i.e., the global vector. The global vector is then processed through a final projection layer and layer normalization to form the output of the module, i.e., a feature sequence containing position context information, i.e., the global feature sequence.

[0051] At the same time, during the entire forward propagation process, the DAI module will monitor and output the deviation of the learnable position parameters from the real physical position. Finally, the dual consideration of position and feature similarity is realized, enabling it to adaptively focus on the most effective dynamic area in the weather field for prediction, thereby significantly improving the prediction accuracy of the wind speed at a fixed point under complex weather phenomena. The DAI module pseudocode is shown in Table 1.

[0052] Table 1 DAI algorithm introduction table

[0053] Step S306, an LSTM-Transformer module is constructed based on the Transformer structure. The Transformer structure used in the LSTM-Transformer module is obtained by adding a linear layer and a normalization layer to the initial Transformer Encoder structure, which can better capture the temporal relationship of the data and strengthen the spatio-temporal continuity of the prediction.

[0054] Step S307, the convolutional long short-term memory network, the dynamic attention interpolation module and the LSTM-Transformer module are connected to obtain an initial model of the offshore wind speed prediction model.

[0055] Step S308, the initial model is trained, verified and tested based on the training set, the verification set and the test set, and the trained, verified and tested initial model is taken as the offshore wind speed prediction model.

[0056] The offshore wind speed prediction model first processes the normalized data through two layers of ConvLSTM, then uses the DAI module to interpolate the wind speed data of the target area grid data set to the target point position, iterates multiple times, feeds the data in each batch into the subsequent LSTM-Transformer module and the subsequent pooling and fully connected layer part, obtains the output of the offshore wind speed prediction model and the original label data for loss function value calculation. The model parameters are updated according to the loss function value to complete the model training. The loss function of the offshore wind speed prediction model updates the weight parameters through the back propagation algorithm. The expression of the loss function is: ; ; ; wherein, is the mean square error, i is the data sample size, i =1,2,…, n , n is a positive integer, y i is the wind speed true value at time t, i pred is the model prediction value, y i+1 is the absolute error of the first derivative of the prediction sequence and the true sequence, which is used to ensure the smoothness of the prediction result in time, grad y i is the wind speed true value at time t+1, is the wind speed prediction value at time t+1, i is the wind speed prediction value at time t, is the loss function. i loss

[0057] ​​​​According to the offshore wind speed prediction model, the input features of the test are fed to obtain the output, and the loss value is calculated with the true label to obtain the final test data loss value, and the effect is visualized, and the experimental results are evaluated using MSE, Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE). The formulas of RMSE and MAE are as follows: ; .

[0058] The comparison experiment visualization and experimental result evaluation of the offshore wind speed prediction model (DAI-LSTM-AT model) and the current commonly used time series prediction model-convolutional neural network (CNN) model, RNN model, LSTM model, bidirectional long short-term memory network (BiLSTM) model, gated recurrent unit (GRU) model, bidirectional gated recurrent unit (BiGRU) model are shown in Table 2.

[0059] Table 2 Comparison and evaluation table of offshore wind speed prediction model and multiple models

[0060] Step S4, input the normalized data into the offshore wind speed prediction model to obtain the offshore wind speed prediction value of the target point position.

[0061] According to Figure 5 and Figure 6 the multi-model comparison chart of No. 1 buoy and No. 2 buoy, for No. 1 buoy, according to the analysis of Figure 5 , the value predicted by the DAI-LSTM-AT model is most consistent with the actual observed value (i.e. True-Value in Figure 5 ), including mean absolute error (MAE), mean square error (MSE), root mean square error (RMSE) and correlation coefficient (R). The performance of DAI-LSTM-AT model on these indicators is 0.672 m / s, 0.778 m 2 / s 2, 0.882 m / s and 0.951. In terms of MAE index, the MAE value of DAI-LSTM-AT model is 0.672 m / s, which is 7.4%~17.6% lower than other models. The MSE value of DAI-LSTM-AT model is 0.778 m 2 / s 2 , which is 16.3%~45.0% lower than other models. In terms of RMSE, the value of DAI-LSTM-AT model is 0.882 m / s, which is 8.5%~25.8% lower than other models. The correlation coefficient (R) of DAI-LSTM-AT model is 0.951, which is higher than BI-LSTM model (0.928), CNN model (0.896), GRU model (0.902), LSTM model (0.931), RNN model (0.923) and BI-GRU model (0.921). For No. 2 buoy, according to the analysis of Figure 6 , the similarity of the predicted value of DAI-LSTM-AT model and the actual observed value (i.e. True-Value in Figure 6 ) is still the highest. In terms of MAE index, the MAE value of DAI-LSTM-AT model is 0.736 m / s, which is 9.3%~19.6% lower than other models. The MSE value of DAI-LSTM-AT model is 0.958 m 2 / s 2 , which is 14.1%~40.3% lower than other models. In terms of RMSE, the value of DAI-LSTM-AT model is 0.979 m / s, which is 8.8%~22.7% lower than other models. The correlation coefficient (R) of DAI-LSTM-AT model is 0.944, which is higher than BI-LSTM model (0.922), CNN model (0.905), GRU model (0.896), LSTM model (0.927), RNN model (0.913) and BI-GRU model (0.912). From Table 2 combining the prediction results of two buoys, DAI-LSTM-AT model is 7.9%~18.1% lower in MAE index and 7.8%~24.2% lower in RMSE index than other models. This comparison evaluation highlights the effectiveness of DAI-LSTM-AT model in the field of wind power prediction, which surpasses other baseline models.

[0062] The offshore wind speed prediction method based on the DAI-LSTM-AT model provided by the application is used for improving the offshore wind speed prediction precision. First, the limitation of the traditional method for considering the feature similarity and the problems of insufficient deep-sea observation data and poor reanalysis data precision are solved, the model based on the stable law is difficult to obtain good results, and the offshore wind speed prediction updates the network weight and position parameter through back propagation, adds the normalization and linear layer on the specified position of the time series model, and improves the prediction precision of the target position in the region while well capturing the surrounding time series features. Secondly, under the deep learning model framework, the loss function is used to update the parameters of the model through back propagation, the ConvLSTM is used to preliminarily extract the space-time features of the data, the DAI module is used to model the target position time series, the LSTM and the modified Transformer structure are used for time series prediction, and finally the features of each time step are further refined through the pooling layer and the full connection layer, so that the application can be widely used for the offshore work ship route point wind speed prediction, the precise prediction of the storm route point wind speed, and can also provide strong support for the implementation regulation of the offshore wind farm, the equipment maintenance and the development of the potential wind power region, and accordingly the manpower and the fund input are reduced.

[0063] The offshore wind speed prediction system based on the DAI-LSTM-AT model is used for realizing the offshore wind speed prediction method based on the DAI-LSTM-AT model, as shown in the following Figure 7 The offshore wind speed prediction system based on the DAI-LSTM-AT model is used for realizing the offshore wind speed prediction method based on the DAI-LSTM-AT model, as shown in the following The data acquisition module is used for acquiring the meteorological grid reanalysis data of a target region and the buoy observation data of a target point position.

[0064] The data processing module is used for performing height conversion and time alignment processing on the buoy observation data according to the meteorological grid reanalysis data, and performing normalization processing on the meteorological grid reanalysis data and the processed buoy observation data to obtain normalized data.

[0065] The model construction module is used for constructing an offshore wind speed prediction model according to a long short-term memory network, a dynamic attention interpolation method and a Transformer structure.

[0066] The prediction module is used for inputting the normalized data into the offshore wind speed prediction model to obtain an offshore wind speed prediction value of the target point position.

[0067] The computer device provided by the application comprises a memory, a processor and a computer program stored in the memory and capable of running on the processor, and the processor executes the computer program to realize the content of the offshore wind speed prediction method based on the DAI-LSTM-AT model.

[0068] The application provides a computer readable storage medium, which has a computer program stored thereon, and the computer program realizes the content of the offshore wind speed prediction method based on the DAI-LSTM-AT model when executed by a processor.

[0069] Finally, it should be noted that the above examples are only used to illustrate the technical method of the present application, but not to limit it. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present application can still be modified or replaced equivalently, and these modifications or equivalent replacements should not make the modified technical method deviate from the spirit and scope of the technical method of the present application.

Claims

1. A method for offshore wind speed prediction based on a DAI-LSTM-AT model, characterized in that, The offshore wind speed prediction method comprises the following steps: S1, obtaining meteorological grid reanalysis data of a target area and buoy observation data of a target point position; S2, performing height conversion and time alignment processing on the buoy observation data according to the meteorological grid reanalysis data, and performing normalization processing on the meteorological grid reanalysis data and the processed buoy observation data to obtain normalized data; S3, constructing an offshore wind speed prediction model according to a long short-term memory network, a dynamic attention interpolation method and a Transformer structure; S4, inputting the normalized data into the offshore wind speed prediction model to obtain an offshore wind speed prediction value of the target point position.

2. The offshore wind speed prediction method based on the DAI-LSTM-AT model according to claim 1, characterized in that, The specific content of constructing the offshore wind speed prediction model according to the long short-term memory network, the dynamic attention interpolation method and the Transformer structure in S3 comprises: obtaining meteorological grid reanalysis historical data of a region, buoy observation historical data of a point position and offshore historical wind speed values of the point position; the point position is in the region; performing height conversion and time alignment processing on the buoy observation historical data according to the meteorological grid reanalysis historical data, and performing normalization processing on the meteorological grid reanalysis historical data and the processed buoy observation historical data to obtain normalized historical data; constructing a training set, a validation set and a test set according to the normalized historical data and the offshore historical wind speed values; constructing a convolution long short-term memory network; the convolution long short-term memory network is obtained by replacing matrix multiplication in the long short-term memory network with convolution operation; constructing a dynamic attention interpolation module based on a learnable self-attention mechanism; constructing an LSTM-Transformer module based on a Transformer structure; connecting the convolution long short-term memory network, the dynamic attention interpolation module and the LSTM-Transformer module to obtain an initial model of the offshore wind speed prediction model; training, validating and testing the initial model based on the training set, the validation set and the test set, and taking the trained, validated and tested initial model as the offshore wind speed prediction model.

3. The offshore wind speed prediction method based on the DAI-LSTM-AT model according to claim 2, characterized in that, The convolution long short-term memory network comprises a forgetting gate, an input gate, an output gate and a memory unit; the multiplication used by the forgetting gate, the input gate and the output gate is convolution multiplication; The convolution long short-term memory network extracts features from the normalized historical data to obtain spatio-temporal features.

4. The offshore wind speed prediction method based on the DAI-LSTM-AT model according to claim 3, characterized in that, The specific content of constructing the dynamic attention interpolation module based on the learnable self-attention mechanism comprises: obtaining latitude and longitude according to the spatio-temporal features; projecting the latitude and longitude through a fully connected encoding layer to obtain a position query vector; mapping the spatio-temporal features through a linear projection layer to obtain spatio-temporal mapping features; the spatio-temporal mapping features have the same dimension as the position query vector; defining an embedding matrix and adding it to the spatio-temporal mapping features element by element to obtain a spatial position fusion feature; performing similarity calculation on the position query vector and the spatial position fusion feature, and performing weighted summation to obtain a global vector; performing projection processing and layer normalization processing on the global vector to obtain a position feature sequence, and obtaining the dynamic attention interpolation module.

5. The offshore wind speed prediction method based on the DAI-LSTM-AT model according to claim 4, characterized in that, The similarity of the position query vector and the spatial position fusion feature is calculated, and a weighted sum is performed to obtain the specific content of the global vector, which includes: The spatial position fusion feature is flattened to obtain a key vector and a value vector. The similarity of the query vector and the key vector is calculated to obtain an attention weight, and the value vector is weighted and summed through the attention weight to obtain the global vector.

6. The offshore wind speed prediction method based on the DAI-LSTM-AT model according to claim 4, characterized in that, The Transformer structure used in the LSTM-Transformer module is obtained by adding a linear layer and a normalization layer to the initial Transformer Encoder structure.

7. The offshore wind speed prediction method based on the DAI-LSTM-AT model according to claim 2, characterized in that, The loss function in the offshore wind speed prediction model updates the weight parameters through the back propagation algorithm; the expression of the loss function is: ； ； ； wherein, is the mean square error, i is the data sample size, i = 1, 2,..., n , n is a positive integer, y i is the i real value of the wind speed at time t, y pred is the predicted value of the model, grad is the absolute error of the first derivative of the predicted sequence and the real sequence, y i+1 is the i real value of the wind speed at time t+1, is the i predicted value of the wind speed at time t+1, is the i predicted value of the wind speed at time t, loss is the loss function.

8. A system for offshore wind speed prediction based on DAI-LSTM-AT model, characterized in that, The offshore wind speed prediction system based on the DAI-LSTM-AT model is used to implement the offshore wind speed prediction method based on the DAI-LSTM-AT model in any one of claims 1-7, comprising: A data acquisition module is configured to acquire meteorological grid reanalysis data of a target area and buoy observation data of a target point position. A data processing module is configured to perform height conversion and time alignment processing on the buoy observation data according to the meteorological grid reanalysis data, and perform normalization processing on the meteorological grid reanalysis data and the processed buoy observation data to obtain normalized data. A model construction module is configured to construct an offshore wind speed prediction model according to a long short-term memory network, a dynamic attention interpolation method, and a Transformer structure. A prediction module is configured to input the normalized data into the offshore wind speed prediction model to obtain an offshore wind speed prediction value of the target point position.

9. A computer device comprising: A memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that the processor executes the computer program to implement the content of the offshore wind speed prediction method based on the DAI-LSTM-AT model in any one of claims 1-7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the content of the offshore wind speed prediction method based on the DAI-LSTM-AT model in any one of claims 1-7.

Images